Update README.md

faq/README.md

@@ -79,7 +79,7 @@ Place this line at the top of your script, below the import statement and before
 
 **Answer:** Our pre-trained 2p DeepInterpolation models were trained on (512, 512) images. It is likely your input movie has odd dimensions causing boundary conditions errors. Essentially the "MaxPooling" layers are downsampling the size of each image, rounding off the dimensions. Then, after upsampling, the concatenation layers that implement the skip connection will receive two inputs of different sizes, as a result of this rounding of image dimensions.  You could either re-trained a network that implement padding (for example see here: https://github.com/AllenInstitute/deepinterpolation/blob/6bec1e7b3cdd83a85b1f95e1ec168210e7a5e18c/deepinterpolation/network_collection.py#L568) OR just pad your input with zeros to match our image size of (512, 512). In practice, our pre-trained networks handle black borders very well as those were present throughout our training data as a consequence of motion correction.
 
-**Question:Which tensorflow version are compatible with DeepInterpolation**
+**Question:Which tensorflow versions are compatible with DeepInterpolation**
 
 **Answer:** Tensorflow is a fast evolving library that can be tightly linked to specific CUDA and python version. Therefore upgrading the Tensorflow package can have significant impacts on your workflow. We encourage any machine Learning effort to be careful when selecting a specific Tensorflow and Python version. We ran tests on the package and found that Tensorflow 2.5 and 2.6 had some package incompatibilities within python 3.7. Tensorflow 2.7 seems to work with DeepInterpolation and Python 3.7 (it passed our integration tests). Notice that our Continuous Integration system is running python 3.7 and we therefore cannot guarantee that everything will work smoothly with more recent versions of python like 3.9. We decided to fix Tensorflow to 2.4.4 within Python 3.7 given these insights. This particular combination could change in the future if needs arise or some critical security issues are discovered. That being said, our Continuous Integration engine is running an additional set of tests on the latest tensorflow version within python 3.7 so as to notice any incoming conflicts.